DEV.co
Open-Source LLM · nvidia

Qwen3.5-397B-A17B-NVFP4

This is NVIDIA's quantized version of Alibaba's Qwen3.5-397B-A17B, a 397-billion-parameter mixture-of-experts language model compressed to FP4 precision for efficient inference. It accepts text, images, and video inputs, supports a 262K token context window, and is optimized for NVIDIA GPUs. The model is permissively licensed (Apache 2.0), ungatED, and ready for immediate commercial and non-commercial deployment. Performance benchmarks show minimal degradation compared to the FP8 baseline across multiple reasoning and coding tasks.

Source: HuggingFace — huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
480.1k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developernvidia
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads480.1k
Likes102
Last updated2026-06-30
Sourcenvidia/Qwen3.5-397B-A17B-NVFP4

What Qwen3.5-397B-A17B-NVFP4 is

Qwen3.5-397B-A17B-NVFP4 is a post-training quantized MoE transformer with 397B total parameters but only 17B activated per token, reducing compute and memory footprint. Quantization was performed using NVIDIA Model Optimizer v0.42.0, applying FP4 precision to weights and activations of linear operators within transformer MoE blocks. The model supports multimodal input (text, RGB images, MP4/WebM video) and generates text output. Calibrated on CNN DailyMail and Nemotron-Post-Training-Dataset-v2. Tested on NVIDIA B200 hardware. Inference engines: SGLang (primary), vLLM. Context length: 262K tokens. Training dataset details (modality, collection, labeling) are undisclosed.

Quickstart

Run Qwen3.5-397B-A17B-NVFP4 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="nvidia/Qwen3.5-397B-A17B-NVFP4")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Large-Scale AI Agent & Chatbot Systems

The 262K context window and MoE sparse activation enable efficient multi-turn conversations and stateful agent reasoning at scale. FP4 quantization reduces serving costs on large clusters.

Code Generation & Technical Problem-Solving

Benchmarks show strong performance on LiveCodeBench (0.843) and SciCode (0.479), making it suitable for code assist tools, automated testing frameworks, and technical documentation generation.

Retrieval-Augmented Generation (RAG) Systems

Long context and quantized inference fit RAG pipelines where external documents are concatenated for grounded generation. Efficient MoE activation reduces latency at scale.

Running & fine-tuning it

**ESTIMATE — Verify with your infrastructure team.** With FP4 quantization and MoE (17B active per token), approximate GPU memory per instance: 80–150 GB for multi-GPU tensor parallelism (4–8 GPUs). Tested on NVIDIA B200; compatible with Blackwell. FP4 weights only; assume ~50–100 GB for model weights + activations + KV cache at typical batch sizes. Exact VRAM depends on batch size, max context length (262K), and tensor parallel configuration. Production serving typically requires 4–8 H100/B200 GPUs.

Not explicitly documented. The model is post-training quantized to FP4; full-parameter fine-tuning would require dequantization, incurring high memory overhead. LoRA/QLoRA fine-tuning feasibility is unknown and vendor documentation does not address it. Contact NVIDIA or test on a small pilot. Standard practice: fine-tune the base Qwen3.5-397B-A17B (unquantized) then quantize, rather than fine-tuning the FP4 version directly.

When to avoid it — and what to weigh

  • CPU-Only or Non-NVIDIA Deployment — Model is optimized for NVIDIA GPUs (tested on B200, Blackwell architecture). CPU inference or non-NVIDIA accelerators are not documented or supported. Deployment outside GPU clusters incurs severe performance penalties.
  • Real-Time, Single-Token Latency-Critical Applications — MoE models with sparse activation and 397B parameters require tensor parallelism across multiple GPUs. Single-GPU inference or strict <50ms latency SLAs are not feasible. Throughput optimized, not latency optimized.
  • Datasets with Heavy Non-English or Specialized Domain Bias — Training data is undisclosed. CNN DailyMail calibration is English-language news. If your use case requires strong performance on low-resource languages, specialized vertical domains (e.g., biomedical NER), or non-text modalities, in-domain evaluation is required.
  • Applications Requiring Guaranteed Absence of Toxic or Biased Output — Model card acknowledges the base model was trained on internet data containing toxic language and societal biases. No filtering or fine-tuning for safety is documented. Requires guardrails and human review for sensitive use cases.

License & commercial use

Apache License 2.0. This is a permissive, OSI-approved license allowing commercial and non-commercial use, modification, and distribution, provided copyright and license text are retained. Model card explicitly states: 'This model is ready for commercial/non-commercial use.'

Apache 2.0 is permissive and explicitly permits commercial use. Model card states 'ready for commercial/non-commercial use.' However: (1) This is NVIDIA's quantized derivative of Alibaba's Qwen3.5-397B-A17B. Verify that Alibaba's base model license is compatible with your use (Alibaba's Qwen models are typically Apache 2.0 or Qwen License; check Qwen/Qwen3.5-397B-A17B directly). (2) Model card recommends developers 'work with their internal model team to ensure this model meets requirements for the relevant industry and use case.' For regulated industries (healthcare, finance, automotive), internal compliance review is mandatory. (3) No SLA, warranty, or indemnification is stated. Use is at your risk.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model is unquantified for embedded adversarial robustness or formal verification. Considerations: (1) Quantization to FP4 may reduce robustness to adversarial perturbations vs. full-precision baseline; no formal analysis provided. (2) Model amplifies biases and toxic content from training data; prompt injection and jailbreaking risks are plausible and not mitigated by the model itself. (3) NVIDIA provides a vulnerability reporting channel (intigriti VDP link). (4) Inference engines (SGLang, vLLM) should be kept current to patch supply-chain or runtime vulnerabilities. (5) No formal security audit or threat model is documented. For regulated use, threat modeling and adversarial testing are required.

Alternatives to consider

Meta Llama 3.1 (70B or 405B)

Permissive Llama 2 License (Apache-compatible for commercial use under certain restrictions). Smaller (70B) for lower cost; larger (405B) for capability parity. Broader ecosystem (vLLM, TGI, Ollama, llama.cpp). No MoE sparsity but simpler quantization landscape.

Alibaba Qwen3.5-397B-A17B (unquantized FP8 or BF16)

Same base model, full precision or higher precision quantization. If VRAM and inference latency are not critical, better model quality. Avoid if cost and scale are primary drivers.

Anthropic Claude (via API)

Closed-source, hosted alternative for text-only tasks. No infrastructure overhead; strong on safety and alignment. Higher per-token cost; not self-hostable; no multimodal (video) input.

Software development agency

Ship Qwen3.5-397B-A17B-NVFP4 with senior software developers

Assess your GPU infrastructure, validate quantization performance on your workload, and configure tensor parallelism with SGLang or vLLM. Start with a small pilot on 4–8 GPUs. Contact us for deployment architecture review or fine-tuning strategy if your use case requires customization.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Qwen3.5-397B-A17B-NVFP4 FAQ

Can I use this model commercially in production?
Yes. Apache 2.0 license explicitly permits commercial use. Model card states 'ready for commercial/non-commercial use.' However, verify compatibility with Alibaba's base model license and ensure your use case complies with any relevant regulations (e.g., export controls, data residency, bias mitigation). Consider internal legal review for high-risk use cases.
How much GPU memory do I need?
ESTIMATE: 80–150 GB for multi-GPU tensor parallelism (4–8 NVIDIA H100/B200 GPUs). FP4 quantization significantly reduces weight size (~50 GB); activations and KV cache scale with batch size and context length (262K max). Test on a pilot deployment to confirm exact requirements for your batch size and serving configuration.
Can I fine-tune this model?
Not explicitly supported in the model card. Full-parameter fine-tuning of the FP4 model is not documented and would require dequantization. Best practice: fine-tune the base unquantized Qwen3.5-397B-A17B, then quantize using NVIDIA Model Optimizer. LoRA/QLoRA feasibility is unknown; contact NVIDIA for guidance.
Does this model support vision/video?
Model card states inputs include 'Text, Image, Video' and formats include 'RGB, MP4/WebM.' Specifics of multimodal handling (e.g., video frame sampling, image resolution) are not detailed. Assume text is primary and vision is auxiliary. Test on representative samples before production deployment.

Custom software development services

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Qwen3.5-397B-A17B-NVFP4 is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Qwen3.5-397B-A17B?

Assess your GPU infrastructure, validate quantization performance on your workload, and configure tensor parallelism with SGLang or vLLM. Start with a small pilot on 4–8 GPUs. Contact us for deployment architecture review or fine-tuning strategy if your use case requires customization.